3D Object Detection Method Based on Supervised Monocular Depth Estimation in Virtual Point Cloud

Abstract

The invention relates to a virtual point cloud three-dimensional target detection method based on supervised monocular depth estimation, comprising: step 1, using laser radar to measure, and collecting scene depth information; step 2, using the data set training sheet obtained in step 1 mesh depth estimation model. The beneficial effects of the present invention are: the present invention directly uses the camera as the main sensing means, avoids the application of expensive sensors such as laser radar in the three-dimensional target detection system, and also directly avoids the joint calibration existing in the multi-sensor sensing method The problem of synchronizing with data further reduces the cost of sensor deployment, reduces the dependence on lidar, helps to reduce the cost of 3D target detection methods, and promotes the application of this technology in various fields. In addition, the algorithm model is deployed to the edge device through offline training and online prediction, which relieves the computing pressure on the device and improves the intelligence level of the edge device.

Description

technical field [0001] The invention belongs to the technical field of target detection, in particular to a virtual point cloud three-dimensional target detection method based on supervised monocular depth estimation. Background technique [0002] Target detection technology is one of the most important tasks in environmental perception, which mainly perceives the position and category of target objects through images. This technology is widely used in many fields such as industry, transportation, aerospace, medicine and so on. The traditional target detection technology is mainly based on two-dimensional detection, and the detection task of the target object is generated by generating a two-dimensional detection frame. In order to further improve the perception level, many research works in recent years have extended the two-dimensional detection frame to the three-dimensional detection frame to obtain a more detailed pose state of the target object. However, since the im...

AI Technical Summary

Problems solved by technology

This method directly uses lidar as the hardware device for environment perception, which is expensive and cannot be applied to general scenarios on a large scale.

The acquisition scheme of lidar and camera requires joint calibration between devices. If there are problems such as position deviation, it needs to be re-calibrated, and the process is relatively complicated.

In addition, this solution also has data synchronization problems. The frequency of the collected images is inconsistent with that of the point cloud data. It needs to be synchronized before the target detection can be performed.

Images